MiniMax H3 Max: free fast AI video from fal. Fal has released MiniMax H3 Max, a post-trained version of MiniMax H3 tuned for speed: 5-second 768p clips with synced audio in under three seconds. Fal has released MiniMax H3 Max, a post-trained version of MiniMax's open-weights H3 video model tuned for speed. It renders a 5-second, 768p clip with synchronized audio in under three seconds, and it landed at the top of the image-to-video arena the week it shipped. Try it: five free clips a day. Sign in to fal and you get five free H3 Max generations every 24 hours, each up to 15 seconds at 768p with synced audio. Open the text-to-video endpoint, write a prompt, and pick an aspect ratio from 21:9 down to 9:16 for vertical shorts. Because a 5-second clip comes back in roughly the time it takes to read this sentence, you can iterate on shot ideas in a single sitting instead of queuing renders and walking away. Why it matters for creators. H3 Max is not a new base model. Fal took the open weights of MiniMax H3 and retrained them for stronger prompt adherence and better aesthetics, then served the result on its own inference stack. The trade is resolution for speed: H3 Max caps at 768p where the base model reaches 2K, but it generates fast enough to feel interactive. For creators who tested audio-synced generation in its ComfyUI H3 sound-sync coverage, this is the same model family with a free, hosted, near-real-time front end. Key details. Base model: Post-trained from MiniMax H3 open weights by fal, developed by MiniMax. Resolution: 480p or 768p (default 1344x768 at 24 FPS), tuned for speed over maximum resolution. Speed: A 5-second 768p clip renders in under three seconds; 15-second clips take about 15 seconds. Free tier: Five 15-second generations every 24 hours with synchronized audio. Pricing: $0.06 per second at 768p, with a 50 percent introductory discount for the first 14 days. Rankings: First for image-to-video on Design Arena and first with audio on the Artificial Analysis video leaderboard. What to do next. Run the same prompt through H3 Max and your current video tool and compare prompt adherence and motion. If the 768p ceiling works for social shorts and rough cuts, the free daily quota makes it a low-cost way to storyboard before committing paid credits on a higher-resolution model for the final render.

fal announced fal Agent, a creative AI partner that orchestrates production-ready workflows across leading image, video, and 3D generative models. The tool maintains context and consistency throughout projects, allowing creatives to move between different models without rebuilding prompts or losing creative decisions. fal Agent automatically selects appropriate models for each task whilst preserving characters, objects, and visual styles across generations. Projects retain memory and references, enabling teams to resume work weeks later with full context intact. The platform integrates with fal's API and CLI, making it accessible for developers building AI-powered creative tools. It runs on fal's inference infrastructure, trusted by over 2.5 million developers. fal Agent is now available in early access at fal.ai/agent, with add-on credits usable across the entire fal platform.

fal adds LoRA training for MiniMax H3, starting with an open-source Realism People LoRA. Custom LoRA training for MiniMax H3 is now live on fal, and the company demonstrated it with Realism People, an open-source LoRA it says pushes the model toward photorealistic humans. * The trainer targets MiniMax H3. It fine-tunes the open-weight model that generates video, image, audio, and text. * Realism People ships open source. fal posted the LoRA weights to Hugging Face, tuned for skin, eyes, and motion. * The trainer covers several input modes. fal points to entry points for text-to-video, image-to-video, first-last-frame, and reference-to-video. * More LoRAs are coming. fal says additional LoRAs are on the way. Realism People targets photorealistic humans. fal trained Realism People to show what the new trainer produces, describing it as a LoRA that pushes H3 toward raw, photorealistic humans with a focus on skin, eyes, and motion. fal posted the weights as an open-source release on Hugging Face, so you can run the LoRA directly or study how it was built before training your own. Because it is an open release rather than a hosted-only feature, the LoRA can be pulled into other workflows built around H3 rather than staying locked to fal's interface. VP Land has also compared reference images versus LoRAs for holding a consistent look across shots. Training on an open-weight video model. A LoRA lets you steer a base model toward a specific look, subject, or style without retraining the whole thing, which is how fal can layer training on top of H3. If the term is new to you, VP Land broke down what a LoRA is for non-technical readers. fal lists trainer entry points spanning text-to-video, image-to-video, first-last-frame, and reference-to-video, so a custom LoRA can attach to more than one generation path. Teams wanting a consistent character or aesthetic across shots can tune H3 on their own material and reuse the result instead of prompting it blind each time. A trainer and reference LoRA make H3 customizable. The MiniMax H3 launch gave creators an open-weight model that unifies multiple generation types. Adding a trainer and a reference LoRA turns that base model into something you can customize, and the open-source Realism People release gives a concrete starting point rather than a closed demo. With fal signaling more LoRAs to follow, the near-term value is a growing library of tunable looks for anyone already generating with H3.